基于多元提升核神经网络的机械故障诊断方法及其特征提取可解释性研究  

Neural Network Driven by Multiple Lifting Kernels for Mechanical Fault Diagnosis and Its Interpretability Research of Feature Extraction

作　　者：袁静 任港星 蒋会明 赵倩 魏臣隽 朱骏 YUAN Jing;REN Gangxing;JIANG Huiming;ZHAO Qian;WEI Chenjun;ZHU Jun(School of Mechanical Engineering,University of Shanghai for Science and Technology,Shanghai 200093;Shanghai Radio Equipment Institute,Shanghai 201109)

机构地区：[1]上海理工大学机械工程学院,上海200093 [2]上海无线电设备研究所,上海201109

出　　处：《机械工程学报》2024年第12期51-64,共14页Journal of Mechanical Engineering

基　　金：国家自然科学基金:(52375111,51975377,52005335,52205113);上海市青年科技英才扬帆计划(21YF1430600)资助项目。

摘　　要：以卷积神经网络为代表的深度学习方法为机械故障诊断大数据分析与处理提供有效工具,但其底层逻辑和物理内涵等“黑盒”问题破解是发展可信、安全、可靠人工智能及机械故障智能诊断方法的一个重要研究方向。提升多小波框架是一个天然多通道卷积过程,基于多小波基函数内积匹配思想可有效提取隐藏于背景噪声下多种故障特征。因此,将提升多小波理论引入卷积神经网络,提出基于多元提升核神经网络的机械故障诊断方法并探讨其底层多通道卷积下故障特征提取机理。首先,该网络以提升多小波框架为底层构架设计自适应提升多小波层,并在提升多小波理论数学约束下构造兼备重要信号处理特性的多元提升核,通过单参数训练高效精准完成多故障特征匹配提取。其次,通过仿真试验研究该网络基于内积匹配原理的物理内涵,探讨训练过程中多元提升核波形演化规律并研究其多通道卷积运行机理、网络映射关联含义等特征提取可解释性问题。最后,试验验证表明该方法对同工况类间差异小、多工况类内差异大特性下行星齿轮箱故障识别表现出优异诊断准确性、稳定性和抗噪性,工程应用表明该方法对高精密天线指向机构微弱和复合故障识别也具备精确诊断能力。Deep learning method represented by convolutional neural network(CNN)provides an effective tool for big data analysis and processing of mechanical fault diagnosis,but the crack at"black box"issue is one of the important research fields of credible,safe and reliable artificial intelligence and its mechanical fault intelligence diagnosis methods.Lifting multiwavelet framework is a natural multichannel convolution process,and could effectively extract multiple fault features hidden in background noise based on the idea of multiwavelet inner product matching.Therefore,this paper introduces the lifting multiwavelet theory into CNN,proposes the neural network driven by multiple lifting kernels for mechanical fault diagnosis,and discusses the fault feature extraction mechanism for its underlying multichannel convolution.Firstly,the network designs an adaptive lifting multiwavelet layer by lifting multiwavelet framework as the underlying architecture.Multiple lifting kernels are constructed with important signal processing characteristics by the mathematical constraints of lifting multiwavelet theory.Multiple fault feature matching and extraction could be accurately and efficiently achieved by training a single parameter.Secondly,the physical connotation of the network based on the inner product matching principle is studied by simulation experiments.The waveform evolution law of multiple lifting kernels in the training process is discussed.The interpretability problems for feature extraction such as the operation mechanism of multichannel convolution and relationship meaning of network mapping are studied.Finally,the experimental case shows that the method has nice diagnostic accuracy,stability and anti-noise for fault identification of planetary gearboxes with small inter-class differences among the same working conditions and large intra-class differences within multiple working conditions.Meanwhile,the engineering application shows that the method also has accurate diagnosis ability for weak and multiple fault identi

关 键 词：卷积神经网络 提升多小波 可解释性 特征提取 

分 类 号：TH17[机械工程—机械制造及自动化]